Avoidance of care: how health-care affordability influenced COVID-19 disease severity and outcomes

Abstract

The COVID-19 crisis revealed several underlying issues in the United States’ health-care system, including issues related to health-care affordability.¹ According to the 2019 National Center for Health Statistics data brief, 14.5% of adults aged 18-64 years were uninsured preceding the pandemic, with the most commonly cited reason being unaffordability.² In an April 2020 poll conducted during the pandemic, 1 in 7 participants reported that they would avoid seeking care if they had symptoms of COVID-19, and 9% reported they would avoid health-care encounters even with a confirmed COVID-19 diagnosis, specifically citing concerns regarding health-care costs.³

When under- or uninsured individuals anticipate costly health-care expenses, they may delay or avoid care, which in turn can manifest as undiagnosed and/or poorly managed chronic diseases. Several studies have confirmed economic disparities in severity and outcomes among persons with cancer, renal and liver diseases, and diabetes, where the under- and uninsured are far more likely to present to a health-care provider with late-stage or uncontrolled conditions.^4,-11 Additionally, individuals without insurance are more likely to skip preventive services and not have a regular source of health care compared with those with coverage.¹² In the wake of the pandemic, these facts are particularly alarming, as many of the comorbid conditions disproportionately experienced across socioeconomic groups were associated with more severe COVID-19 illness and a higher risk of death.¹³

In this study, we examined the association between health-care coverage status and the presenting severity of illness among individuals acutely hospitalized with symptomatic COVID-19. We hypothesized that the under- and uninsured population would present with a higher COVID-19 disease severity, suggesting possible consequences of missed routine care visits, poorly managed chronic disease, and/or a choice to delay or avoid care.

Methods

In this retrospective analysis, we aimed to identify associations between health-care coverage and presenting severity of illness among nonelderly individuals acutely hospitalized with COVID-19 disease. The study population included patients admitted to New York University Langone Health (NYULH; New York, New York) between March 1, 2020, and June 30, 2020. Inclusion required a positive COVID-19 polymerase chain reaction (PCR) test. Patients were excluded if they were admitted to the labor and delivery unit for childbirth; this population was universally tested but rarely demonstrated symptoms, and did not represent people with symptomatic disease seeking attention for COVID-19. To enrich the analysis with persons most likely to be without health-care coverage, we analyzed only those aged 18-64 years, since individuals outside of this age bracket are likely to have Medicare coverage. Lastly, we excluded patients with home addresses outside of the 5 boroughs of New York City and Nassau County, since our access to pre- and postadmission health data might have been limited. This study was approved by the NYU Grossman School of Medicine Institutional Review Board.

Measures

The primary outcome was severity of COVID-19 illness assessed within the first 24 hours of admission, expressed using the modified sequential organ failure assessment (mSOFA) score. The mSOFA score is a measure of cumulative organ failure and is a predictor of intensive care unit mortality.¹⁴ The mSOFA quantifies organ-dysfunction severity on a 0- to 4-point scale for the central nervous, cardiovascular, respiratory, and renal organ systems, and with a 0- versus 3-point score for the hepatic organ system. Guidelines issued by the New York State Task Force on Life and the Law and the New York State Department of Health, formulated in anticipation of resource shortages during a pandemic, utilize the SOFA score to support clinical decision-making for ventilator allocation.¹⁵ During the COVID-19 surge, our hospital system did not need to activate these guidelines.

Herein, we classify mSOFA severity of disease at presentation into the following 3 categories:

• Asymptomatic-to-mild severity of illness: mSOFA score 0-3

• Moderate-to-severe severity of illness: mSOFA score 4-7

• Critical severity of illness: mSOFA score >7

Individuals who died within 24 hours of admission were assigned an mSOFA score greater than 7 (n = 4; 0.2%). mSOFA score organ system components were quantified using reported laboratory values and clinical flow sheets as detailed below.

Central nervous system

Traditionally, the central nervous system component of the mSOFA assesses brain dysfunction using the Glasgow Coma Scale.¹⁶ The Glasgow Coma Scale measures impaired consciousness by gauging motor, verbal, and eye-opening responsiveness,¹⁶ and the score can be confounded by sedatives used for intubation, which was common among severe COVID-19 cases. Because we found that all COVID-19 patients were alert upon admission, for this study patients were assumed to have a normal (0) central nervous system score.

Liver

The liver component of the mSOFA measures the presence of sclera icterus or jaundice. For our analysis, liver dysfunction was indicated by serum bilirubin levels outside of the laboratory normal range within 24 hours of admission. Individuals without measured bilirubin levels were assumed to have no clinical suspicion of liver dysfunction and were given a score of 0.

Renal system

Creatinine levels, assessed within 24 hours of admission, were used to measure renal function. Missing laboratory values were imputed (number of missing observations (n_m) = 30).

Respiratory system

The respiratory component of the mSOFA score was calculated from the oxygen-saturation (S_pO2)/inspired-oxygen-fraction (F_iO2) ratio, calculated within 24 hours of admission; a lower S_pO2/F_iO2 value indicates worse respiratory function. To ensure the validity of the mSOFA score, S_pO2 measurements less than 60% (n = 16) were excluded, because keystroke error was suspected. F_iO2 levels not expressly recorded were estimated using recorded oxygen delivery methods. F_iO2 levels reported to be less than 20% were excluded (n = 66). Finally, excluded and missing data (total S_pO2  n_m = 28; total F_iO2  n_m = 265) were imputed.

Cardiovascular system

The cardiovascular component of the mSOFA score measures circulatory failure with the assessment of blood pressure and use of vasoactive agents used to treat hypotension. The first indicator of hypotension is an observed mean arterial blood pressure less than 70 mm Hg. The first mean arterial pressure measured within 24 hours of admission was extracted from flow sheets. Measurements less than 40 mm Hg were excluded (n_m = 4). Severity of hypotension was also indicated by the administration of continuous vasoactive infusions of dobutamine, dopamine, epinephrine, norepinephrine, phenylephrine, or vasopressin. While phenylephrine and vasopressin were not included as vasoactive agents in the original derivation of the mSOFA score, these drugs were commonly used for circulatory failure during the COVID-19 pandemic. Due to the rapidly changing clinical situations in the most severe cases of COVID-19 pneumonia, the absence of automated recording of vasoactive drip doses, and the incredible clinical pressure placed on providers during the first wave of COVID-19 cases in our hospital, we felt that flow sheets would not accurately depict discrete drip rates. Instead, we elected to assign the maximum mSOFA cardiovascular score to any patient receiving a vasoactive agent. After manual chart review, we found this estimation to reflect true dosing patterns well.

Clinical covariates

Other patient-level characteristics included in these analyses included age, sex, race, body mass index (weight (kg)/height (m)²), the presence and provider of health-care insurance coverage, Elixhauser Agency for Healthcare Quality and Research (AHQR) Index score, admission calendar week and time of day, and distance between a patient’s home address and the admitting hospital. Patients were categorized by type of health insurance coverage into 4 groups: privately insured, Medicaid, Medicare, or uninsured. For this study, recipients of the Health Resources and Services Administration COVID-19 Uninsured Testing and Treatment Fund were considered uninsured. Preexisting medical conditions were determined from International Classification of Diseases, Tenth Revision, diagnosis codes associated with encounters predating the COVID-19 hospitalization and diagnoses recorded upon admission to the acute-care setting. Prior outpatient and inpatient encounters included any telemedicine, inpatient, and outpatient encounters at an NYULH-associated facility. In addition to these patient-level variables, analyses incorporated census-tract–level population count, race proportion, citizenship rate, educational attainment, median income, Gini index of income inequality, unemployment rate, and health-care coverage rate. The Gini index is a measure of resource distribution (eg, income) in a population, with a value ranging from 0 (ie, a perfectly equal distribution of resources) to 1 (ie, perfect inequality, where 1 individual receives all resources).¹⁷ Patient-level data were extracted from the medical record via the NYULH COVID-19 De-Identified Clinical Database. Census-tract–level data were obtained using the R package tidycensus.¹⁸

Statistical analysis

Missing data included variables pertaining to the mSOFA score components (as detailed in the “Measures” section above) and body mass index (n_m = 113). Rather than impute mSOFA score directly, we calculated missing scores from imputed component values. Under the assumption of data being missing at random, we performed random forest multiple imputations to fill in missing values. Independent variables that did not contain missingness were used as predictors in the imputation model. Ten imputed data sets were created and subsequently combined, or pooled, into a single data set for univariate and multivariate analyses. The “mice” function in the mice R package,¹⁹ the “merge_imputations” function in the sjmisc R package,²⁰ and the “with” function in the mitools R package²¹ were used to conduct multiple imputation, to merge imputed data sets, and to pool results of models fitted on imputed data, respectively.

Descriptive statistics are expressed as median values and interquartile ranges (IQRs) for continuous variables and as numbers and percentages for categorical variables; as appropriate, comparisons were performed using the Kruskal-Wallis rank sum test and the χ² test. The Spearman correlation coefficient (r_s) was computed for census-tract–level variables.

Covariates considered in the model included patient-level variables with a univariate P value less than or equal to .10 when comparing mSOFA scores, and county-level variables with a correlation coefficient of r_s < 0.70. Summary statistics were generated using the “tbl_summary” function in the gtsummary R package,²³ and the ACL model was built using the “bracl” function in the R package brglm2.²⁴

Analyses were performed on both a data set comprised of complete (ie, nonimputed) data (n = 1697) and a data set including imputed data (n = 2108). Because results of these two analyses were so similar, results from the complete-case analyses are presented in the main body of this paper, and results of the imputed analyses are provided in supplemental tables.

Results

Between March 1 and June 30, 2020, a total of 2108 PCR-positive COVID-19 patients aged 18-64 years were admitted to one of 3 NYU Langone Health acute-care hospitals for reasons other than childbirth. Indeed, during this period, more than 73.7% of inpatient hospitalizations were expressly for treatment of COVID-19 pneumonia.

Table 1 demonstrates patient-level characteristics stratified by type of health insurance coverage. Medicare recipients represented the smallest proportion of the overall patient population (10.1%), while nearly half of the analyzed population (48.1%) was privately insured. Medicare recipients appeared to have the worst overall baseline health: 62% had an Elixhauser AHRQ Index score greater than or equal to 5, as compared with 34%, 39%, and 29% of the private, Medicaid, and uninsured populations, respectively (P <.001). This is consistent with New York State eligibility rules for Medicare, under which Medicare is granted to persons below age 65 years with chronic conditions,²⁵ and is consistent with the higher rate of pre–COVID-19 outpatient visits in this population (64%) as compared with other groups (39%, 36%, and 21% in the private, Medicaid, and uninsured populations, respectively; P <.001). The uninsured population represented the second-smallest subgroup (18.3%). This group was also the youngest, with a median age of 47 years; was majority male (79%); and was overwhelmingly of Hispanic ethnicity (74%).

Univariate analyses supported a link between poor baseline health, under- or uninsured health-care status, and more severe disease at presentation. In the complete-case analysis (n = 1794), Medicare recipients had the highest rate of a moderate-to-severe mSOFA score; 34% of those covered by Medicare presented with an mSOFA score of 4-7, as opposed to 13%-17% of the other coverage groups (P <.001) (Table 1). The uninsured had the highest rate of critical mSOFA scores on admission: 10% presented with an mSOFA score greater than 7, as compared with 5%-6% in the other groups (P =.03). In Table 2, we also observed a general trend of higher rates of preexisting conditions associated with higher mSOFA scores at presentation. The greatest disparities existed in the rate of preexisting renal failure between those with asymptomatic-to-mild (6.7% in mSOFA group 0-3) and moderate-to-severe (33% in mSOFA group 4-7; P <.001) disease. Similar results were seen in the imputed case analyses, shown in Table S1 (n = 2108).

Though the group with critical disease at presentation was only a small subset of the overall patient population, the difference in hospital experience in this group versus those with less severe disease was profound. Nearly half (47%) of those presenting with a critical mSOFA score died in the hospital (Table 2)—8 times the death rate among those presenting with asymptomatic-to-mild disease (5%; P <.001). Those presenting with a critical mSOFA score also had substantially longer lengths of stay than those with an asymptomatic-to-mild disease presentation. Among hospital encounters that did not end in death, the median length of stay in the critical population was over 2.5 weeks (the IQR ranged from 8 days to almost 1.5 months), as compared with a median length of stay of 5 (IQR, of 3-8) days in the asymptomatic-to-mild population (P <.001). As an additional reflection of the relationship between presenting stage of disease and hospital experience, Table S2 shows intubation events that occurred at any time during the acute hospitalization. Seventy-nine percent of patients presenting with critical disease were intubated and ventilated during the course of their disease, versus 29% and 8% of the patients presenting with moderate-to-severe and asymptomatic disease, respectively (P <.001).

Table 3 shows the results of the ACL model used to fit the log odds of (1) the adjacent pair moderate-to-severe and asymptomatic-to-mild presentations and (2) the adjacent pair critical and moderate-to-severe presentations using our complete-case population (n = 1697). Similar ACL model results for the imputed data set are reported in Table S3. Consistent with the univariate analysis, the ACL model demonstrated a relationship between health insurance status and disease severity at presentation. Compared with the privately insured, uninsured patients had twice the odds of presenting with a critical mSOFA score as opposed to moderate-to-severe disease (odds ratio (OR) = 2.08, P =.03).

In addition, the ACL model identified several factors associated with increased odds of presenting with moderate-to-severe (mSOFA 4-7) versus asymptomatic (mSOFA 0-3) disease, including male sex (OR = 2.07, P <.001), increased age (OR = 1.02, P =.002), and having preexisting renal failure or presenting with renal failure (OR = 5.66, P <.001). Interestingly, patients with renal failure demonstrated decreased odds of being admitted with critical disease as opposed to moderate-to-severe disease (OR = 0.46, P =.03), contrary to the direction of the relationship in the other adjacent pair. There was little meaningful association between time of day of presentation and disease severity; however, later calendar week of admission was associated with decreased odds of presenting with critical versus moderate-to-severe disease (OR = 0.92, P =.05). This makes logical sense according to lessons learned during the months of the COVID surge and is also reflected in a lower death rate during the later months of the peak.

Discussion

In this retrospective study, we investigated whether a more severe disease presentation (measured using mSOFA score) disproportionately affected the under- or uninsured among nonelderly, PCR-positive COVID-19 patients admitted to one of 3 NYU Langone Health acute-care hospitals. Payor type—a proxy for health-care affordability—was of prime interest, as we hypothesized that apprehension regarding health-care costs and/or forgone preventative services may have fueled a hesitancy toward seeking care and/or worsened disease severity during the peak of the COVID-19 pandemic.

Univariate analyses indeed supported a link between poor baseline health, insurance status, and presenting disease severity. Upon admission, the highest rate of critical disease (ie, mSOFA score >7) was seen in the uninsured population, and the nonelderly Medicare recipients had the highest rate of moderate-to-severe disease (ie, mSOFA score 4-7).

Consistent with the univariate analyses, multivariable models further supported relationships between insurance status and the most severe disease. With all other covariates remaining fixed, the uninsured had twice the odds of presenting with critical disease as the privately insured (OR = 2.08, P = 0.03). There was a lesser degree of association between the most severe disease and other patient demographic and health factors. A preexisting history of renal failure increased the odds of presenting with moderate-to-severe mSOFA (versus asymptomatic disease; OR = 5.66, P <.001), but contrary to conventional hypothesis, it decreased the odds of presenting with the worst-severity illness (versus moderate-to-severe disease; OR = 0.46, P =.03). The reason for this is not clear but may be related to perceived or actual health-care accessibility. Increased age and being male were only associated with increased odds of presenting with moderate-to-severe versus asymptomatic disease, not with the highest disease severity. Taken together, these results suggest that health insurance, rather than baseline health, was significantly related to the most severe disease presentation. This raises the question as to whether the severity of disease at presentation might have been mitigated by reducing the burden of health-care costs for those most vulnerable.

This question is worth asking, for multiple reasons. First, on the acute, micro-level, the most severe COVID-19 presentations came with significant health costs to the individual. In our population, a “critical” presenting mSOFA score (<7%) was relatively rare, but it was associated with a significant absolute and comparative risk of death (47%). As the COVID-19 experience evolves, we are observing chronic compounding of health-care affordability issues. Though there has been much progress in the prevention of COVID-19–related severe illness, hospitalization, and death, persons who are and have previously been infected with this disease are still at risk of suffering long-term side effects,²⁶ including but not limited to respiratory, cardiac, neurological, and digestive symptoms. In addition to comorbid conditions and increased age, most at risk are those who suffered severe COVID-19 illness and those affected by health inequities.²⁷ Second, on the macro-level, the most severe COVID-19 presentations imposed significant resource-related and financial costs on the health-care system. The hospital length of stay in our “critical” population was 4 times that for persons presenting with mild or asymptomatic disease (2.5 weeks vs 5 days), and the rate of intubation and ventilator use was 10 times higher than among those presenting with asymptomatic disease (79% vs 8%). To illustrate the financial impact, the hospitalization and treatment costs for complex inpatient encounters for COVID-19 (eg, encounters requiring ventilation or admission to the intensive care unit) were estimated nationally to range between $70 000 and $208 000 across payor types, while the median cost of care for noncomplex cases ranged between $25 000 and $54 000.²⁸

The strain that the pandemic placed on hospital resources—human and physical—is harder to calculate, and is more profound. Scarcity of medical resources in the United States is only likely to occur in situations of extraordinary utilization, as was demonstrated by the COVID-19 pandemic experience. Protective measures (social distancing, masking, and so on) engaged in during the acute phase of the COVID-19 outbreak were meant to “flatten the [epidemic] curve” of disease prevalence, such that cases would not exceed health-care system capacity. But the strain on health-care resources might also have been mitigated by reduced disease severity, had it been within the influence of the health-care system; indeed, multiple medical treatments for mitigating COVID severity were urgently tested while current, proven treatments and vaccines were still in development. In this context it is imperative to consider what influence accessible, preventative health care and reduced health-care costs might have had on the ultimate impact of the pandemic.

The disproportionate number of uninsured persons presenting with critical COVID-19 may suggest a delay in seeking treatment or the consequences of unmanaged comorbidity. We were not able to expressly identify or distinguish between these factors in this study, and as such we cannot define the specific roles that health-care affordability and/or access played in the impact of COVID-19 on our hospital system and the individuals within. However, paired with survey data collected during the pandemic,³ our findings are disturbingly consistent with individuals’ attestations of predicted behavior and rationale. Additional studies should explore the role of Medicare and illness severity among nonelderly persons who have qualified for Medicare based on terminal illness (such as end-stage renal failure) or because of chronic disability and limited income. Both preexisting comorbidity and delayed care are plausible causes for the more severe presentation in the Medicare population; an analysis of payor types among those with a documented history of advanced-stage diseases could be conducted to better understand the specific influence of health-care affordability on illness. To the extent that avoidance of care or uncontrolled comorbidity worsened COVID-19 outcomes, addressing both is an opportunity to prevent avoidable deaths and to mitigate long-term disease sequelae, while saving financial and physical resources on both the national and provider levels during the next event of extraordinary health disturbance and health-care utilization.

This study had several limitations. First, it was impractical to distinguish the privately insured population from the underinsured, that is, individuals with higher copays, deductibles, and other out-of-pocket expenses. Those opting for low-cost/low-coverage plans may yet experience financial pressures that affect behavior.²⁹ Based on our coverage definition, underinsured patients were possibly absorbed into the privately insured subgroup, but this would have narrowed the observed differences in outcomes related to disease severity—that is, biased the results toward the null hypothesis, which was that insurance status was not associated with disease severity. Second, preexisting conditions were only captured for individuals with prior encounters in our hospital system, or persons who reported their conditions to an admitting nurse; true preexisting comorbidity rates may have been underrepresented. Third, 19.5% of the components necessary to calculate the primary dependent variable were missing. The largest source of missingness was missing or erroneous F_iO2 flow sheet entries (F_iO2 < 20%, n_m = 265). We can draw no conclusions about the nature of this patient population, but we tested whether this missingness biased conclusions, and results based on the complete data set (n = 1794) and the imputed data set (n = 2108) were equivalent with respect to our primary findings. Finally, generalization of results to other populations should take into account that this study cohort was obtained from a single health-care enterprise, consisting of 3 acute-care hospitals, in the densely populated New York metropolitan area, which experienced a surge of disease less than 3 months after the first documented US case, when no proven therapy or vaccine existed. Moreover, patient demographic characteristics, testing and hospital resources, population-level epidemiologic curves, and the patient-level natural history of diseases represented in this study may not reflect the experiences of people at other times or in other parts of the country or globe. Nonetheless, it is reasonable to anticipate that another pandemic of a novel respiratory virus for which there is no established treatment or vaccine may result in an experience similar to that represented in these pages. For instance, the 2022 global monkeypox outbreak demonstrated an unlikelihood of COVID-19’s being a singular episode.

This experience should serve as a motivation to mitigate the burden of future health-care disasters—both acute and chronic—via policy action. As an excellent example of policy positively influencing public health outcomes, early estimates suggest that mass vaccinations achieved by the universal COVID-19 vaccine rollout reduced numbers of new cases by millions and prevented several thousand COVID-related hospitalizations and deaths.^30,31 However, further measures should be considered, as it was evident during the 2020 pandemic that human and physical health-care resources are exhaustible.^32,-36 We hope our study results are taken into strong consideration in the planning for allocation and promotion of claims reimbursement in the inevitable event of future population-level health-care demands. In the more immediate setting, we hope our study results serve as a cost- and resource-saving argument for more affordable health-care access for those most in need.

Acknowledgments

We express our sincere gratitude for the contribution of the NYU Langone Medical Center Information Technology COVID-19 Clinical Data Mart and acknowledge the diligent efforts made in the creation of the study data set.

Supplementary material

Supplementary material is available at American Journal of Epidemiology online.

Funding

None declared.

Conflict of interest

There are no conflicts of interest to report.

Data availability

Due to privacy and legal restrictions, the data utilized in this study are not publicly available.

References